Shared memory + knowledge-curation skills; seed a cross-role knowledge layer

[arozumenko]

Per-role memory is local and role-scoped, so a fact one role pays for is invisible to every other role — and to the same role on another machine. Teams rediscover, expensively, what a teammate already diagnosed correctly weeks earlier. This adds the missing second layer and the path into it.

memory → one shared skill

It was bundle-local in feature-development and test-automation, and the two copies had drifted. The divergence was accidental — test-automation's copy had corrected overclaims about snapshot.md being host-generated — so that is the version promoted. Both local copies deleted; registered in skills.json.

memory → an Obsidian-grade vault, not "just notes"

Entries now carry:

• aliases — so [[service token]] resolves to service_token_rotation.md, and search finds a note by the words you'd actually type rather than the slug
• tags on closed cross-cutting axes (type/, area/, status/) — a tag used once filters nothing; relationships are links, not tags
• created / updated — where updated means when the fact was re-verified, not when the file was touched
• ## sections so [[note#Section]] works, and Related: links — a note with no links is one nobody rediscovers

Unknown keys are tolerated, never stripped: a human edits this vault in Obsidian too. Plus a routing decision tree (board / knowledge / curated / daily / nowhere), upkeep rules, and an audit snippet for unindexed entries and dead index lines.

Verified before touching frontmatter that nothing parses it — bin/init.mjs's ensureMemoryIndexLine only writes an index line. The old "don't add extra keys" warning referred to a snapshot generator that doesn't exist in this repo.

knowledge-curation — new shared skill

Owns .agents/knowledge/ and the path from memory into it:

• promote — four admission tests: cross-role · verified · durable · costly to rediscover
• sweep — end of session/mission: "would another role have needed this today?"
• audit — unindexed notes, dead links, notes overdue re-verification
• retire — correct or delete; never leave a half-true note

Admission is deliberately narrow, because an unverified claim in a committed layer is worse than silence — it is trusted.

Wired into all four bundles

Both skills go into each bundle's team-wide skills, so manual-qa and product-management get them too. Neither previously shipped a memory skill at all.

Scout seeds the layer

New Step 2.5 in seeding-a-project: scaffold .agents/knowledge/ with a charter and per-folder READMEs stating what belongs where, seed it from what onboarding actually verified, and record the contract in CLAUDE.md/AGENTS.md — outside any <!-- BUNDLE --> markers, so regeneration can't clobber it.

Applied to both copies rather than promoting the skill: unlike memory, seeding-a-project's divergence is intentional (test-automation's carries TMS-specific content), so consolidating would have lost content. Four other bundle-local skills diverge the same way — playwright-testing, reproducing-issues, test-automation-workflow, test-case-analysis — left alone for the same reason.

instructions.md ×4 carries the two-layer contract into each generated AGENTS.md.

Ideas taken from the existing obsidian-vault skill

It's aimed at personal-assistant, but several patterns generalise and are reused here: a documented frontmatter schema with per-type extensions, a decision tree for "where does this go?", lint/audit as a first-class operation, and — the most valuable — "unknown fields are tolerated, never blocking".

Verification

node bin/validate-bundles.mjs green on all 4 bundles; every bundle's team-wide skills resolve against the registry; no bundle-local memory directories remain.

🤖 Generated with Claude Code